Capturing the dynamics of learning and forgetting through adaptive retrieval practice
Across virtually all levels and topics of education, students need to learn facts. The SlimStampen algorithm, developed at the UG, aims to optimize the process of fact learning. PhD student Thomas Wilschut explains how the algorithm works, what it tells us, and how future work (including his own) can improve it.
Every university course requires learning factual knowledge. Whether it concerns the names of nerves or cell types in medicine, dates, and events in history or brain structure functions in psychology, in order to pass exams it is a necessity to know your facts and terminology. Unfortunately, learning them takes time and effort. At our faculty, a frequently used approach to facilitate memorizing such facts is the SlimStampen computerized adaptive learning system, which presents facts or glossary items to the student for practice.
While students are working their way through their SlimStampen assignments, they are largely unaware of the algorithm that is running in the background. This algorithm consists of a set of mathematical equations that aim to describe how we remember and forget information over time. When a student is faced with a question like ‘what is the name of the brain structure that is responsible for maintaining balance, coordination and fine motor skills?’, then their response and response time are used as input to these equations. The algorithm then allows us to get a better idea of what is happening in the student’s mind. We can estimate how well a student has memorized a fact, predict at which point in time a fact will be forgotten, optimize item repetition schedules for more efficient learning, and even generate individual memory measures. Interested in how it’s done? Keep reading.
A model of human memory
At the core of the adaptive learning system lies the assumption that all pieces of information, once encountered, have a certain associated memory strength or activation. Without additional practice, this activation declines over time. The decline of memory activation after an encounter, is what we call the rate of forgetting. The SlimStampen algorithm assumes that this rate of forgetting can be changed with practice. If we only study the names of brain structures in the context of a course, and never speak or think about that information after the exam, it will be forgotten rather quickly. However, if we keep encountering the information, the rate at which the memory activation declines after each encounter goes down.
Optimizing item repetition schedules to improve learning
Using this rate of forgetting, it is possible to estimate the point in time at which the memory activation of an item drops below the so-called forgetting threshold: the point in time at which the fact cannot be retrieved anymore. The SlimStampen memory model posits that repeating a practice item just before it is forgotten is optimal. This is based on literature that suggests that when retrieval attempts are effortful but successful, learning is optimally efficient. Repeating the item sooner is unnecessary, as the student already knows the answer, and trying to retrieve the item would be a waste of time. Repeating the item after the student has forgotten the answer means that the retrieval attempt comes too late: Unsuccessful retrieval attempts may result in motivational issues and have a high associated time cost, as corrective feedback needs to be presented after each incorrect answer. Repeating items just before they are estimated to be forgotten, has proven to be a successful learning approach for a range of materials and settings: students typically learn about 10% more items in the same study time compared to using non- or less-adaptive learning protocols.
Apart from accuracy and response times, we can use other metrics to infer memory strength. Some of these come to expression through spoken language.
Using behavioral parameters to estimate memory strength
Crucial to the success of adaptive learning systems is their ability to estimate and predict future memory performance. When a student practices facts with SlimStampen, we can use behavioral parameters that are recorded during learning to estimate the rate of forgetting for individual students and individual facts. There are several ways to measure forgetting from observable behavior, including the response itself, characteristics in speech, and even pupil size. The SlimStampen algorithm mainly relies on accuracy and response times. If you give a correct response quickly, then there is a good chance that the item is highly active in memory and that the rate of forgetting of this item is low. If a student needs a long time to think about an answer, the memory representation is weak, and the rate of forgetting is higher. Think about the following example: imagine that you are first asked about the capital of France; next, try to recall the name of the capital of Albania. Notice how long it took you to come up with either answer. Answers that are strongly represented in memory come to mind quickly, where answers that are more difficult to recall take longer.
Apart from accuracy and response times, we can use other metrics to infer memory strength. Some of these come to expression through spoken language. In my PhD project, I have shown that it is possible to use the prosodic information in speech – in other words, the way in which the student gives a spoken response – to estimate the extent to which the student has successfully memorized a response. More specifically, we found that students that raise their voice while giving a response are likely to be uncertain of their answer, and increased loudness and speaking speed are generally predictors of high memory strength (and accurate answers). Interestingly, even if no human experimenter is present (i.e., when students are interacting with an app like SlimStampen) participants use such prosodic cues. In an extended analysis, we found that we could use the information present in current-repetition spoken responses to retrieval practice questions to improve predictions of future retrieval performance success. Overall, detecting prosodic features in speech may be a relatively straightforward and computationally inexpensive way to better predict which questions a student should practice with at which point during the learning session.
Inclusive and effective: adaptive learning with specific learning disabilities
No learner is unique: learners differ in capabilities, learning speed and preferences. One of the main advantages of adaptive learning systems is that they can personalize learning sessions — tailoring them towards the needs of specific (groups of) learners — to an extent that is not possible in traditional classroom learning. In the two remaining years of my PhD, I want to explore if we can combine the insights in adaptive learning and speech analysis techniques to aid learners with specific learning disabilities, such as developmental language disorder or dyslexia. First, speech-based learning may be beneficial for these learners, because it does not rely on the processing of written text as much as traditional, typing-based learning methods. The project also aims to examine the possibility of using speech characteristics to estimate an individual memory metric for each learner, in order to facilitate quick personalization of the learning session.
As described above, the SlimStampen algorithm tries to estimate the rate of forgetting for each fact and for each student separately. Yet, we can also calculate an average or median rate of forgetting for a student across different questions, and use it as an individual memory metric. Studies show that the rate of forgetting estimated over multiple sessions is highly correlated amongst individuals, even when diverse sets of learning materials are used. Similarly, recent work shows that the estimated rate of forgetting is associated with neurophysiological activity at rest (EEG and fMRI). In other words, participants’ patterns in brain activity, when they were not doing any specific task, were predictors of later estimated rate of forgetting in a learning task. This underlines the idea that the rate of forgetting can be characterized as a trait-like marker of memory. In the current project, we want to test if we can further refine these individual memory measures by taking prosodic speech information into account. Using the learners’ answers to practice questions, their prosodic speech features, and reaction time distributions, we ultimately aim to recognize if a learner belongs to a specific learner group, and adjust the difficulty of the learning session accordingly.
Overall, what appears to be a simple set of responses to retrieval practice questions can prove to be a window into memory dynamics. The time we need to retrieve an answer, the accuracy of the response, or the exact way in which we speak to a learning system, can give valuable insights into the way we learn, remember, and forget. Subsequently, we can use this information to create a learning session that is optimally efficient for each individual learner — including those that are typically underrepresented in education.
References
More about the science behind SlimStampen: https://www.memorylab.nl/en/wetenschap/
Academic publications
Sense, F., Behrens, F., Meijer, R. R., & van Rijn, H. (2016). An individual’s rate of forgetting is stable over time but differs across materials. Topics in cognitive science, 8(1), 305-321.
Van Rijn, H., Dalenberg, J. R., Borst, J. P., & Sprenger, S. A. (2012). Pupil dilation co-varies with memory strength of individual traces in a delayed response paired-associate task. PLoS One, 7(12), e51134.
Van Rijn, H., van Maanen, L., & van Woudenberg, M. (2009). Passing the test: Improving learning gains by balancing spacing and testing effects. In Proceedings of the 9th international conference of cognitive modeling (Vol. 2, No. 1, pp. 7-6).
Wilschut, T., Sense, F., Scharenborg, O., & van Rijn, H. (2023). Improving Adaptive Learning Models Using Prosodic Speech Features. In International Conference on Artificial Intelligence in Education (pp. 255-266). Cham: Springer Nature Switzerland.
Zhou, P., Sense, F., van Rijn, H., & Stocco, A. (2021). Reflections of idiographic long-term memory characteristics in resting-state neuroimaging data. Cognition, 212, 104660.
Thomas has completed a Bachelor program in Psychology followed by a Research Master program in Behavioural and Cognitive Neurosciences at the University of Groningen. Now, he is in his PhD program in the Experimental Psychology unit at the university of Groningen. His PhD constitutes a collaboration between the UG and the Groningen-based company MemoryLab, where he tries to find ways to optimize the benefits of studying with SlimStampen, and make it easily accessible to all learners.
